Beam Download (Latest): BEAM_release1.6.1
Beam Download (Official Github): BEAM_release_1.5.1
Bear Encoder Download: BearEncoder (Usage: java -jar <encoder.jar> <input.db> <output.fb> [where .db is a fasta with primary sequence and dotbracket for every RNA])
Bear Encoder Download (2nd version, in case the first throws strange errors): BearEncoder.new
RNALfold to BEAM-friendly DotBracket format python script: lfoldToDb.py. (Usage: python lfoldToDb.py <RNALfold output>)

Typical pipeline for treating local structure predictions as suboptimals


Since BEAM relies on secondary structure predictions, it is our duty to provide you with ways to deal with the inherent error rate of these tools.
It can happen that one of your datasets is composed by very long RNAs (typically > 300nt). Folding it globally with standard methods may result in very inaccurate structures, and BEAM will have a very hard time in looking for a motif which may not have been predicted in most of the RNAs (even if we can manage up to 80% of noise, see the paper for more information).
As a workaround, one can tell BEAM to consider many different predictions for any specific RNA, by appending to the ID a progressive counter in this form "$<counter>", resulting in something like this:

Input file:
>ID1$1
PrimarySequence1
Prediction1
>ID1$2
PrimarySequence1
Prediction2
...
>ID2$1
PrimarySequence2
Prediction1

And so on.
Now, if BEAM is launched with the additional option -m T, it will treat these structures as mutual exclusive instances for the same RNA.
(-Warning: technical terms here-)
This is done by adding another possible perturbation during the algorithm that swaps the current selected instance.
(-End of technical terms-)
This was originally thought to include different predictions of the same whole structure (e.g. Suboptimals), but it can be exploited without shenanigans to deal with smaller local predictions.
The only information you do not want to lose are the start positions of the local predictions.
You should do this by yourself but we have just what it takes:
With the script that you can download on the top of the page you can convert the output of RNALfold (with standard parameters) in a Fasta file ready to be launched on the webserver (remember to add the -m T option in the Advanced > Add Command Area) or ready to be converted by the BearEncoder for command line use.
All the information about the starting points of the local predictions is found in the IDs in the following format:

>ID|start:<start_0-based>$<prediction_counter>

In this way you can trace back a signal to its original position for every RNA.
Simpy download the "RNALfold to BEAM-friendly DotBracket format python script", and feed it with the RNALfold output that you previously generated (python lfoldToDb.py <RNALfold output>) and you are ready to submit a run to the BEAM Web Service!
MP